Codex 现在支持可复用的云环境，并能扫描仓库以查找安全漏洞。Agents API 接入了 Computer Use 功能，同时 OpenAI 正在推出 Decisions API 和 Ultrafast——一种价格较高的新速度层级。

OpenAI 在旧金山举行的 DevDay 2026 开发者大会上宣布了这些更新。开发人员还将获得 revamped（重新设计）的 Codex CLI、ChatGPT 桌面应用中的代码审查视图，以及 OpenAI 与亚马逊合作运营的代理平台的新技术基础。

可复用环境让团队能更快地启动 Codex 任务

据 OpenAI 称，Codex 现在可以在你自己的电脑上、手机上或从任何设备通过云端运行。OpenAI 早在五月就通过智能手机上的 ChatGPT 应用增加了控制功能。新的亮点是可复用的开发环境，旨在更快地启动任务。团队可以使用它们共享具有已批准设置和权限的配置。云端的 Codex 适用于 Plus、Pro、Business、Enterprise、Edu 和 Healthcare 计划。

用户可以通过语音向 Codex 下达任务并大声指挥其工作。| 图片来源：OpenAI
OpenAI 还重新设计了 Codex CLI。用户现在可以通过语音启动和指挥任务。新的“/agents”视图旨在简化委托过程，并能同时显示多个任务。OpenAI 还改进了提示词编辑、会话恢复以及与 Git worktrees 的工作方式，并清理了终端界面。新的 CLI 适用于所有计划。

Codex 在桌面上审查代码，并在云端搜寻漏洞

ChatGPT 桌面应用获得了专用于跨多个项目更改的代码审查视图。开发人员可以阅读摘要、查看差异，并在向 GitHub pull requests 或 GitLab merge requests 留下反馈之前询问 Codex 关于潜在问题的意见。通过自动审查，Codex 会在开发人员离开电脑时在云端进行初步检查。该功能适用于所有计划。

在新的代码审查视图中，Codex 会对更改发表评论并根据请求应用修复。| 图片来源：OpenAI
Codex Security Cloud 面向安全团队。它可按需或按计划扫描整个 GitHub 仓库，并在新提交进入时持续检查。Codex 会调查其发现的内容，过滤掉重复项，并在云端准备修复方案，即使笔记本电脑处于关闭状态。七月，OpenAI 发布了作为开源工具的 Codex Security CLI，该工具可从命令行和 CI/CD 管道中检查仓库。

云版本的用户还可以访问 Daybreak Blue 模型，而无需单独申请 Daybreak。Daybreak Blue 是 OpenAI 网络安全计划的防御层级，专为恶意软件分析等任务构建。在此之前，它需要身份验证等其他操作。Codex Security Cloud 适用于桌面和网页端的 Pro、Business、Enterprise 和 Edu 计划。

Codex Security Cloud 跟踪其发现的每一个漏洞，从发现到修复。| 图片来源：OpenAI
基于 API 构建的代理现在可以自行操作软件

Agents API 自 9 月 11 日起处于公开测试阶段，并运行在与 Codex 和 ChatGPT 相同的基础设施上。新的补充功能是 Computer Use，它允许代理自主操作软件以完成任务。OpenAI 自四月起在 macOS 上的 Codex 和 ChatGPT 中提供此功能，自五月起在 Windows 11 上提供。

OpenAI 表示，该 API 现在还包括 Codex 的多代理功能、Tool Search 以及 Context Compaction（自动压缩长上下文）。

Agents API 处理代理的控制循环，而命令和文件在沙箱中运行。| 图片来源：OpenAI
AWS 和 OpenAI 四月宣布 Bedrock Managed Agents 为有限预览。该平台现在承担 Agents API 的核心功能，可针对 AWS 进行定制，并直接连接到 AWS 资源。这意味着开发人员可以构建完全在 AWS 上运行的 OpenAI 代理。

OpenAI 加入对快速决策模型的推动

通过 Decisions API，OpenAI 正进入所谓的 System One 模型领域。9月中旬，Diogo Almeida 的初创公司 TypeSafe AI 推出了此类模型之一 Jev，这位前 OpenAI 研究人员的发布在开发者中引发了一波炒作热潮。将更强的模型作为编排器与快速的决策模型配对看起来很有前景。例如，一个 Minecraft 智能体使用 GPT-6 Astra 规划目标和航点，而 Jev 则选择挖掘、制作或战斗等具体动作。

对于 Decisions API，OpenAI 使用了其低成本模型 GPT-6 Luna 的一个版本。开发者通过限定的一组预定义答案来定义问题，以文本或图像形式传入上下文，并获取返回的答案。他们可以用它来分类内容、路由请求或决定智能体的下一步操作。该 API 以有限预览的形式推出，预计将在未来几天内广泛可用。

OpenAI 表示，Decisions API 的决策速度比通过常规 API 使用的 GPT-6 Luna 快十倍。| 图片来源：OpenAI

超快的速度意味着六倍的价格

根据 OpenAI 的说法，新的 Ultrafast 高级层将 Codex 中的令牌生成速度提高了多达八倍，达到每秒 300 个令牌。在 API 中，速度提高了多达六倍。Ultrafast 的 API 定价是标准费率的六倍。对于 GPT-6 Astra，这意味着每百万输入令牌 60 美元，每百万输出令牌 300 美元。

OpenAI 以双倍价格提供 Astra 的 Fast 模式，速度为原来的 2.5 倍。GPT-5.3-Codex-Spark 要快得多，在 Cerebras 硬件上每秒可生成超过 1,000 个令牌，但它是一个较小的模型，基准测试分数较弱。

GPT-6 Astra Ultrafast 现已在 API 中提供。它也面向企业客户和订阅新版 Pro 500 计划的 ChatGPT Work 和 Codex 用户。该计划售价 500 美元，旨在提供任何订阅中最高的使用额度，尽管 OpenAI 并未明确说明具体有多高。现有的 200 美元 Pro 计划现在在 ChatGPT Work 和 Codex 中仅包含十倍于 Plus 的使用量，而不是之前的二十倍。ChatGPT 中的 GPT-6 Pro 消息数量也减半，从每周 200 条减少到 100 条。现有客户在过渡期内保留其当前限制，并获得一次性信用额度。

新的 GPT-6.1 Sol 在 API 中的价格与其前身相同，即每百万输入令牌 2 美元，每百万输出令牌 10 美元。预计 Ultrafast 版本将在未来几天内推出。

没有炒作的 AI 新闻 – 由人工策划

订阅 THE DECODER，享受无广告阅读、每周 AI 通讯、我们每年六次的独家“AI Radar”前沿报告、完整档案访问权限以及我们的评论区访问权限。

立即订阅

Codex now gets reusable cloud environments and can scan repositories for security holes. The Agents API picks up Computer Use, and OpenAI is adding a Decisions API and Ultrafast, a pricey new speed tier.

OpenAI announced the updates at its DevDay 2026 developer conference in San Francisco. Developers also get a revamped Codex CLI, a code review view in the ChatGPT desktop app, and a new technical foundation for the agent platform OpenAI runs with Amazon.

Reusable environments let teams get Codex tasks running faster

According to OpenAI, Codex now works on your own computer, from your phone, or in the cloud from any device. OpenAI already added control through the ChatGPT app on smartphones in May. What's new are reusable development environments that are supposed to get tasks started faster. Teams can use them to share one configuration with approved settings and permissions. Codex in the cloud is available on Plus, Pro, Business, Enterprise, Edu, and Healthcare plans.

Users can give Codex tasks by voice and steer its work out loud. | Image: OpenAI
OpenAI has also reworked the Codex CLI. Users can now start and steer tasks by voice. A new "/agents" view is meant to make delegating easier and shows several tasks at once. OpenAI also improved prompt editing, session resuming, and work with Git worktrees, and cleaned up the terminal interface. The new CLI is available on all plans.

Codex reviews code on the desktop and hunts for vulnerabilities in the cloud

The ChatGPT desktop app gets its own code review view for changes across multiple projects. Developers can read summaries, go through diffs, and ask Codex about possible problems before leaving feedback on GitHub pull requests or GitLab merge requests. With automatic reviews, Codex does a first pass in the cloud while the developer is away from their computer. The feature is available on all plans.

In the new code review view, Codex comments on changes and applies fixes on request. | Image: OpenAI
Codex Security Cloud is aimed at security teams. It scans entire GitHub repositories on demand or on a schedule and keeps checking new commits as they come in. Codex investigates what it finds, filters out duplicates, and prepares fixes in the cloud, even when the laptop is closed. In July, OpenAI released the Codex Security CLI as an open-source tool that checks repositories from the command line and in CI/CD pipelines.

Users of the cloud version also get access to the Daybreak Blue models without having to apply for Daybreak separately. Daybreak Blue is the defensive tier of OpenAI's cybersecurity program , built for tasks like malware analysis. Until now, it required identity verification, among other things. Codex Security Cloud is available for Pro, Business, Enterprise, and Edu on desktop and web.

Codex Security Cloud tracks every vulnerability it finds from discovery to fix. | Image: OpenAI
Agents built on the API can now operate software themselves

The Agents API has been in public beta since September 11 and runs on the same infrastructure as Codex and ChatGPT. The new addition is Computer Use, which lets agents operate software on their own to get tasks done. OpenAI has offered the feature in Codex and ChatGPT on macOS since April and on Windows 11 since May.

OpenAI says the API now also includes Codex's multi-agent features, Tool Search, and Context Compaction, which automatically condenses long contexts.

The Agents API handles the agent's control loop while commands and files run in a sandbox. | Bild: OpenAI
AWS and OpenAI announced Bedrock Managed Agents as a limited preview in April . The platform now takes on the core features of the Agents API, can be customized for AWS, and connects directly to AWS resources. That means developers can build OpenAI agents that run entirely on AWS.

OpenAI joins the push for fast decision models

With the Decisions API, OpenAI is moving into the field of so-called System One models. In mid-September, Diogo Almeida's startup TypeSafe AI introduced one such model called Jev , and the former OpenAI researcher's release set off a wave of hype among developers. Pairing a stronger model as an orchestrator with a fast decision model looks promising. One Minecraft agent , for example, has GPT-6 Astra plan goals and waypoints while Jev picks individual actions like mining, crafting, or fighting.

For the Decisions API, OpenAI uses a version of GPT-6 Luna , its low-cost model. Developers define questions with a limited set of predefined answers, pass in context as text or images, and get an answer back. They can use it to classify content, route requests, or decide an agent's next step. The API launches as a limited preview, with broad availability expected in the coming days.

OpenAI says the Decisions API makes decisions ten times faster than GPT-6 Luna does through the regular API. | Bild: OpenAI
Ultrafast speed comes at six times the price

According to OpenAI, the new Ultrafast premium tier speeds up token generation in Codex by up to eight times, to 300 tokens per second. In the API, it's up to six times faster. API pricing for Ultrafast is six times the standard rate. For GPT-6 Astra , that works out to $60 per million input tokens and $300 per million output tokens.

OpenAI offers Astra's Fast mode at double the price for 2.5 times the speed. GPT-5.3-Codex-Spark was much faster, hitting more than 1,000 tokens per second on Cerebras hardware, but it was a smaller model with weaker benchmark scores.

GPT-6 Astra Ultrafast is available now in the API. It's also in ChatGPT Work and Codex for Enterprise customers and subscribers to the new Pro 500 plan. That plan costs $500 and is supposed to offer the highest usage allowance of any subscription, though OpenAI doesn't say exactly how high it is. The existing $200 Pro plan will now include only ten times Plus usage in ChatGPT Work and Codex instead of twenty times. GPT-6 Pro messages in chat also get cut in half, from 200 to 100 per week. Existing customers keep their current limits for a transition period and get a one-time credit.

The new GPT-6.1 Sol costs the same in the API as its predecessor, $2 per million input tokens and $10 per million output tokens. An Ultrafast version is expected in the coming days.

AI News Without the Hype – Curated by Humans

Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section.

Subscribe now